Instruction optimizers: MIPROv2 and GEPA

Demo optimizers choose examples. Instruction optimizers rewrite the signature's instructions, using a model to propose new wording and your metric to pick the best.

So far every optimizer left Triage.instructions, "Sort a support ticket for an online shop.", as you wrote it. The instruction optimizers change it. They need a model that can write good instructions, which is why the course shows them without running them: the stand-in's keyword rules cannot propose wording, and a hosted model's proposals change every run, so no output here could be checked.

MIPROv2

import dspy

dspy.configure(lm=dspy.LM("openai/gpt-4o-mini"))

optimizer = dspy.MIPROv2(metric=exact, auto="light", num_threads=8)
optimized = optimizer.compile(dspy.Predict(Triage), trainset=trainset, valset=devset)
print(optimized.signature.instructions)

MIPROv2 works in three stages. It bootstraps demo sets, like lesson 17. It asks the model to propose candidate instructions, showing it a summary of the data, the program's code and example runs. Then it searches combinations of instruction and demo set with Bayesian optimization, evaluating each on batches of the validation set. auto="light", "medium" or "heavy" sets the budget. It needs pip install "dspy[optuna]".

GEPA

def exact_with_feedback(example, prediction, trace=None, pred_name=None, pred_trace=None):
    ok = example.category == prediction.category
    feedback = "Correct." if ok else f"Expected {example.category}, got {prediction.category}."
    return dspy.Prediction(score=float(ok), feedback=feedback)


optimizer = dspy.GEPA(metric=exact_with_feedback, auto="light", reflection_lm=dspy.LM("openai/gpt-4o"))
optimized = optimizer.compile(dspy.Predict(Triage), trainset=trainset, valset=devset)

GEPA reads text, not only scores. Its metric returns a Prediction with score and feedback. A reflection_lm, usually a stronger model, reads failed runs with their feedback and writes improved instructions, and GEPA keeps a population of the best variants. Feedback like "Expected billing: money back is always billing" is information a score cannot carry.

Which one